Remote Sensing of Environment 202 (2017) 98–112

Contents lists available at ScienceDirect

Remote Sensing of Environment

journal homepage: www.elsevier.com/locate/rse

Satellite-based water use dynamics using historical Landsat data

(1984–2014) in the southwestern United States
Gabriel B. Senay a,⁎, Matthew Schauer b, MacKenzie Friedrichs c, Naga Manohar Velpuri d, Ramesh K. Singh d
a
U.S. Geological Survey (USGS), Earth Resources Observation and Science (EROS) Center, North Central Climate Science Center, Fort Collins, CO, USA
b
Innovate! Inc., Contractor to the USGS EROS Center, Sioux Falls, SD, USA
c
SGT Inc., Contractor to the USGS EROS Center, Sioux Falls, SD, USA
d
ASRC InuTeq LLC, Contractor to the USGS EROS Center, Sioux Falls, SD, USA

a r t i c l e i n f o a b s t r a c t

Article history: Remote sensing-based field-scale evapotranspiration (ET) maps are useful for characterizing water use patterns
Received 31 July 2016 and assessing crop performance. The relative impact of climate variability and water management decisions are
Received in revised form 14 April 2017 better studied and quantified using historical data that are derived using a set of consistent datasets and meth-
Accepted 8 May 2017
odology. Historical (1984–2014) Landsat-based ET maps were generated for major irrigation districts in Califor-
Available online 18 May 2017
nia, i.e., Palo Verde and eight other sub-basins in parts of the middle and lower Central Valley. A total of 3396
Keywords:
Landsat images were processed using the Operational Simplified Surface Energy Balance (SSEBop) model that in-
Evapotranspiration tegrates weather and remotely sensed images to estimate monthly and annual ET within the study sites over the
ET modeling 31 years. Model output evaluation and validation using gridded-flux data and water balance ET approaches indi-
Historical ET cated relatively good correspondence (R2 up to 0.88, root mean square error as low as 14 mm/month) between
Landsat SSEBop ET and validation datasets. In a pairwise comparison, annual variability of agro-hydrologic parameters of
SSEBop model actual evapotranspiration (ETa), land surface temperature (Ts), and runoff (Q) were found to be more variable
Thermal remote sensing than their corresponding climatic counterparts of atmospheric water demand (ETo), air temperature (Ta), and
Water use
precipitation (P), revealing process differences between regional climatic drivers and localized agro-hydrologic
responses. However, only Ta showed a consistent increase (up to 1.2 K) over study sites during the 31 years,
whereas other climate variables such as ETo and P showed a generally neutral trend. This study demonstrates a
useful application of “Big Data” science where large volumes of historical Landsat and weather datasets were
used to quantify and understand the relative importance of water management and climate variability in crop
water use dynamics in regards to the linkages among water management decisions, hydrologic processes and
economic transactions. Irrigation district-wide ETa estimates were used to compute historical crop water use vol-
umes and monetary equivalents of water savings for the Palo Verde Irrigation District (PVID). During the peak
crop fallowing year in PVID, the water saved reached a maximum of ~107,200 acre-feet in 2011 with an estimat-
ed monetary payout value of $20.5 million. A significant decreasing trend in actual ET despite an increasing atmo-
spheric demand in PVID highlights the role of management decisions in affecting local hydrologic processes. This
study has importance for planning water resource allocation, managing water rights, sustaining agricultural pro-
duction, and quantifying impacts of climate and land use/land cover changes on water resources. With increased
computational efficiency, similar studies can be conducted in other parts of the world to help policy and decision
makers understand and quantify various aspects of water resources management.
Published by Elsevier Inc. This is an open access article under the CC BY license (http://creativecommons.org/
licenses/by/4.0/).

1. Introduction
Estimates of historical water use, particularly from agricultural
fields, are crucial for planning water resource allocation, managing
water rights, sustaining agricultural production, and quantifying the
⁎ Corresponding author at: U.S. Geological Survey, Earth Resources Observation and
Science Center, North Central Climate Science Center, Fort Collins, CO 80523-1499, USA.
E-mail address: senay@usgs.gov (G.B. Senay).
impacts of climate change and land use/land cover changes on water re-
sources over space and time.
Local and regional hydrologic processes are impacted by the water
lost during evapotranspiration (ET) due to soil-plant-atmosphere inter-
actions. The effect of ET on hydrology and climate particularly in arid/
semi-arid areas with intensive agriculture, such as the Central Valley
in California, shows wide ranging impacts (Felton 1979; Jin et al.
2012; Lo and Famiglietti 2013; Lobell et al. 2009; Shelton 1987;
Sorooshian et al. 2014). Modeling results have shown that irrigation-in-
duced ET exhibits a long-term decreasing trend in the western United

http://dx.doi.org/10.1016/j.rse.2017.05.005
0034-4257/Published by Elsevier Inc. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
 G.B. Senay et al. / Remote Sensing of Environment 202 (2017) 98–112
States (Sorooshian et al. 2014). Based on studies in several major irrigat-
ed regions of the world, Lobell et al. (2009) found substantial regional
differences in the magnitude of irrigation-induced cooling depending
upon differences in extent of the irrigated area, differences in the simu-
lated soil moisture for the control simulation (without irrigation), and
the nature of cloud response to irrigation.
Remotely sensed images are widely used to quantify daily and sea-
sonal ET estimates over large areas in river basins (Anderson et al.
2011; Bastiaanssen et al. 2014; Senay et al. 2016; Singh et al. 2014).
Thermal infrared (TIR) sensors available on Landsat, Moderate Resolu-
tion Imaging Spectroradiometer (MODIS), and other satellites are vital
to capturing variations of land surface temperature (Ts). Ts provides
valuable information about the surface and subsurface moisture status
required for estimating ET and detecting the onset and severity of
drought (Anderson et al. 2011). Since 1972, the Landsat series of satel-
lites has been collecting global images and the Landsat archive is the
world's longest continuously acquired collection of remotely sensed
data. With the launch of Landsat 4 in 1982, Landsat TIR images have en-
abled scientists and modelers to use energy balance models to estimate
ET using spatially explicit Ts derived from the thermal band.
The rich archive of moderately high-resolution Landsat imagery
combined with a simplified energy balance model can reliably quantify
evapotranspiration in irrigated fields (Senay et al. 2013; Senay et al.
2016; Singh et al. 2014; Velpuri et al. 2013). This approach is not only
suitable for scientific investigations of hydrologic processes, but also
provides an advantageous and efficient tool for water management
agencies and farmers to make decisions about water budgeting and irri-
gation scheduling. Researchers have used Landsat time series for under-
standing the surface water extent dynamics at the subcontinental scale
(Mueller et al. 2016; Tulbure et al. 2016). The time is ripe for “Big Data”
remote sensing – the analysis of millions of freely available Landsat im-
ages – to understand historical crop water use dynamics and the links
between water resources and land management decisions.
The U.S. Geological Survey (USGS) has played a critical role in utiliz-
ing and promoting remote sensing data in water resource management.
In 1981, the USGS established an experimental project to assess the pos-
sible and practical use of remote sensing to estimate ET as an approxi-
mation of consumptive use of water in the lower Colorado River Basin
(Raymond and Rezin 1989). In recent years, the USGS started using re-
motely sensed images in implementing the national water availability
and use assessment program (Alley et al. 2013; Michelsen et al. 2016).
The approach for estimating ET using remote sensing data has evolved
from a simple crop coefficient-based method (Raymond and Rezin
1989) to an energy balance-based method using the latest Landsat 8
data (Senay et al. 2016). Remotely sensed images have been used in
two well-documented irrigated regions in the United States, specifically
the Palo Verde Irrigation District (PVID) (Chatterjee et al. 2012;
Elhaddad et al. 2011; Raymond and Rezin 1989) and the Central Valley
(Anderson et al. 2012; Fisher et al. 2007; Marshall et al. 2016; Semmens
et al. 2015) in California. One of the main advantages of this ET time se-
ries is the higher spatial resolution of Landsat imagery and the relatively
long record. These studies (and many more) have provided valuable in-
formation on hydrologic regimes within the study site using satellite
data, but to our knowledge, no study has been carried out to estimate
the historical ET (N 30 years) across these regions using a consistent ap-
proach based on remote sensing data.
The main objective of this study is to quantify and characterize the
spatio-temporal dynamics of historical (1984–2014) ET using Landsat
for the Palo Verde Irrigation District along the lower Colorado River,
and for irrigation districts in California's Central Valley. Furthermore,
seven agro-climatic/hydrologic variables were evaluated for their
spatio-temporal dynamics. Climatic variables refer to those that are
mainly a result of large-scale interactions between solar radiation and
the atmosphere such as atmospheric demand (ETo), daily maximum
air temperature (Ta) and precipitation (P). On the other hand, agro-hy-
drologic variables refer to those parameters that are mainly influenced
99
by land surface processes such as soil moisture and vegetation condi-
tions and operate at a local scale: actual ET (ETa), land surface tempera-
ture (Ts), runoff (Q) and Normalized Difference Vegetation Index
(NDVI).
2. Methodology
2.1. Study sites
2.1.1. Palo Verde Irrigation District
The Palo Verde Irrigation District (PVID) has an area of about
490 km2 in Riverside and Imperial Counties (California) and La Paz
County (Arizona) and is located along the Colorado River bordering Cal-
ifornia and Arizona (Fig. 1). We chose to use hydrological unit codes
(HUC) at the HUC8 level to delineate sub-basin areas from watershed
boundaries (https://water.usgs.gov/GIS/huc.html). In this study, the
PVID boundary we used is the subset of the Imperial Reservoir HUC8
boundary. The predominant soil within the PVID is sandy loam, and
mean annual precipitation is b100 mm. Hot, long growing seasons,
mild winters, and a water supply from the Colorado River allow for mul-
tiple crops to be grown and harvested throughout the year. Alfalfa is the
most dominant crop grown within the PVID among other crops such as
cotton, wheat, barley, maize, vegetables, melons, and citrus. Detailed re-
cords of water supply and consumptive use in the lower Colorado River
Basin are provided by the U.S. Bureau of Reclamation (http://www.usbr.
gov/lc/region/g4000/wtracct.html, accessed on 22 Feb, 2016).
2.1.2. Central Valley
The middle to southern portion of the Central Valley in California,
between the Sierra Nevada Mountains and the Coast Ranges, is one of
the most productive agricultural areas in the United States (Fig. 1). Av-
erage annual precipitation ranges from about 400 mm in the northwest-
ern part of the valley to b 100 mm in the southeast. The Central Valley
climate is characterized by hot, dry summers and mild, rainy winters.
Snowpack on the Sierra Nevada is a major source for the supply of
water within the valley. As the largest single agricultural area in Califor-
nia, this region leads in almost all categories of agricultural production
— total acreage, cultivated acreage, irrigated acreage, water use, and
value of production. Crops grown in the valley include grapes, cotton, al-
monds, citrus, alfalfa, and a diverse range of vegetables. In this study, we
chose eight HUC8 sub-basins (Fig. 1) classified into two groups as Mid-
dle Central Valley (Rock Creek, Lower San Joaquin, Middle San Joaquin
and, Fresno River) and Lower Central Valley (Upper Kaweah, Upper
Tule, Upper Deer and, Upper Poso) sub-basins.
2.2. Input datasets
This study focused on the use of Landsat 5, Landsat 7, and Landsat 8,
which all share the same global orbit (U.S. Geological Survey, 2015). We
downloaded 3396 total Landsat images including 1776 images covering
the PVID (Paths 38–39, Row 37) and 1620 images covering the Middle
and Lower Central Valley sub-basins in southern California (Path/Row
42/35, 43/34).
Table 1 shows the number of Landsat images processed for each of
the 4 path/rows. About 58% of images were from Landsat 5, 38% from
Landsat 7, and 4% from Landsat 8. The number of images used per
year ranged from a minimum of 6 (1984) with one satellite in orbit to
a maximum of 45 (2004) when using two satellites together. The time
interval between used images also varied depending on cloud cover
and the number of satellites in the orbit. In general, there is an 8-day re-
peat cycle during years with two satellites (1999–2011; 2013–2014)
and increasing to a 16-day repeat cycle during years with only one sat-
ellite (1984–1998; 2012).
The thermal band (Band 6 in Landsat 5/7 and Band 10 in Landsat 8)
was used to compute Ts and the NDVI was computed from top-of-atmo-
sphere reflectance in red and near-infrared bands. We used the Fmask
100 G.B. Senay et al. / Remote Sensing of Environment 202 (2017) 98–112

Fig. 1. Map of study areas including the Palo Verde Irrigation District (in the Imperial Reservoir sub-basin) and the eight other HUC8 sub-basins in the Middle and Lower Central Valley.
Rectangular outlines indicate the four Landsat path/rows used in this study.

(Function of Mask) version 3.2 algorithm to detect clouds and cloud
shadows for all satellite types (Zhu et al. 2015).
Other model input data included a gridded (4-km) daily reference
evapotranspiration (ETo) dataset from GridMET (Abatzoglou 2013). Val-
idation of gridded ETo data using automated weather stations across the
western United States has shown it to be comparable to interpolated
station observations (Abatzoglou 2013). Gridded daily maximum air
temperature (Ta) was obtained from the TopoWx dataset, which is
available at 30 arc-second (~800 m) spatial resolution for the contermi-
nous United States since 1948. TopoWx is a collection of gridded esti-
mates of daily minimum and maximum topo-climatic air temperature
that are based on historical daily station observations, digital elevation
model (DEM) variables, atmospheric reanalysis data and MODIS land
surface temperature (monthly average of 2002–2012) (Oyler et al.
2015).
2.3. Other hydrologic datasets
Independent Slopes Model) precipitation datasets (Daly et al. 2007;
Daly et al. 2000). Mean monthly 4 km gridded runoff data for each
study site was processed from HUC8 runoff data by combining historical
flow data from stream gauges, their drainage basins, and the boundaries
of the HUCs (Jian et al. 2008). In order to validate our model results, we
compared our ET estimates to global estimates of land evapotranspira-
tion from 1984 to 2011 from the Max Planck Institute (MPI) in Germa-
ny. These monthly MPI ET estimates are compiled using a machine
learning approach to upscale eddy covariance measurements
(FLUXNET), meteorological data, climate data, and the fraction of
absorbed photosynthetic active radiation data (Jung et al. 2009; Jung
et al. 2010). MPI ET data are available at ~50-km spatial resolution on
a monthly time scale and provide reasonable estimates and longest
time-series of ET estimates when compared to other ET products
(Jung et al. 2009; Velpuri et al. 2013) and these estimates have been
used for validating remote sensing ET results (Senay et al. 2014;
Velpuri et al. 2013).

Other hydrologic datasets included gridded monthly total precipita- 2.4. Crop acreage and fallowing in PVID
tion for the study regions for the period 1984–2014, extracted from
4 km monthly PRISM (Parameter-elevation Relationships on Farmed crop acreages and estimates on fallowing for the Palo Verde
Irrigation District were collected from the PVID Annual Water Depart-
ment Crop Reports provided by PVID Watermaster Bert Bell (personal
Table 1 communication via http://pvid.org/contact.html, last accessed: 02/14/
Distribution of 3396 Landsat images used in the study by sensor type and path/rows (P/R).
2017). The Watermaster provided reports from 1988 (the earliest re-
38/37a 39/37a 42/35a 43/34a Total Imagery dates cord) through 2014. The actual farmed crop acreage from 1988 to
Landsat 5 512 510 450 480 1952 03/1984–01/2013 2014 in the PVID averaged 78,108 acres (31,609 ha), with a maximum
Landsat 7 337 341 312 303 1293 04/1999–12/2014 acreage of 91,990 acres (37,227 ha) in 2000 and reached a minimum
Landsat 8 38 38 38 37 151 02/2013–12/2014 of 47,734 acres (19,317 ha) in 2010. These acreage totals are impacted
Total 887 889 800 820 3396 by a water management agreement between PVID and the Metropoli-
a
P/R: indicate the Landsat Path/Row numbers. tan Water District (MWD) of Southern California to leave a fraction of
 G.B. Senay et al. / Remote Sensing of Environment 202 (2017) 98–112
PVID fields fallow so water can be transferred to meet other demands
outside of PVID (http://pvid.org/PVID-MWD.html, last accessed 7/22/
2016). As a result of this MWD/PVID Fallow Field program, cropped
area in PVID reduced from 88,867 acres (35,963 ha) in 2004 to
62,391 acres (25,249 ha) in 2005 when the program began and
remained below 70,000 acres (28,328 ha) until 2013.
2.5. Water use statistics
The U.S. Bureau of Reclamation (USBR) provides records of water di-
version, returns, and consumptive water use in the lower Colorado
River Basin in both monthly and yearly totals. These reports are pub-
lished annually since 1964 as “Colorado River Accounting and Water
Use Report: Arizona, California, and Nevada” (http://www.usbr.gov/lc/
region/g4000/wtracct.html#decree, last accessed: 02/14/2017). We
collected data (henceforth USBR data) from 1984 to 2014 from these re-
ports for evaluation of ET model results to compare and contrast the es-
timated water use within the PVID study site, which is part of the lower
Colorado River Basin.
2.6. Land cover/land use dataset
In order to isolate irrigated agricultural lands in our analysis of the
final ET products, we used the National Land Cover Database 2011
(NLCD 2011) from the Multi-Resolution Land Characteristics Consor-
tium (Homer et al. 2015). This dataset is classified into several major
land cover types such as water, wetlands, developed spaces, forested
areas, and planted/cultivated areas, among others. We reclassified the
NLCD 2011 to identify planted/cultivated areas (classes 81 and 82) for
use as a mask to extract statistics on ET and other variables for our anal-
ysis of historical trends. The PVID boundary we used in this study covers
areas of California and Arizona (see Fig. 1). However, to enable compar-
ison with reports obtained from the PVID and USBR, we have excluded
farm lands from Arizona.
2.7. Modeling framework
2.7.1. The SSEBop modeling approach
The Operational Simplified Surface Energy Balance (SSEBop) model
computes daily total actual evapotranspiration (ETa) as a product of ET
fractions generated from land surface temperature and reference ET.
In this study, we used the same method to process Landsat images for
ETa as presented in Senay et al. (2016) and Singh et al. (2014). The
SSEBop model does not solve all energy balance terms explicitly; in-
stead of solving ET as a residual, SSEBop directly estimates pixel-based
ETa by calculating an ET fraction (ETf) (0 to 1) from Ts using two key pa-
rameters, i.e., (1) Ta for establishing a maximum ET limit (cold bound-
ary) and (2) a pre-defined temperature difference (dT) between bare-
dry-surface and canopy-level air temperature. Then, ETf is multiplied
by ETo to obtain ETa in mm/day.
The SSEBop model defines limiting (minimum and maximum ET)
conditions based on daily maximum Ta and daily-average net radiation
under a clear-sky assumption (Senay et al. 2013; Senay et al. 2016). Al-
though a brief overview of the SSEBop model is presented here, readers
can refer to Senay et al. (2013) for full details on the model parameters.
The most important assumption of the model is that the surface energy
balance process is mostly driven by available net radiation (Rn), and a
decline in ET due to water stress and other factors can be quantified
by differences in land surface temperature (Senay et al. 2013).
With such simplification, ETa can be estimated using Eq. (1) as a frac-
tion of the ETo.
ET a ¼ ET f  k ∗ ET o ð1Þ
where ETo is the grass reference ET for a location, k is the coefficient to
scale the ETo to the maximum level ET for an aerodynamically rougher
101
crop such as alfalfa (in this study, we used k as 1.25). The ET fraction
(ETf) is calculated using Eq. (2):
T h −T s ðTc þ dT Þ−T s Ts−Tc
ET f ¼ ¼ ¼ 1− ð2Þ
dT dT dT
where Ts (K) is the satellite-observed land surface temperature of the
pixel; Th (K) is the estimated Ts at the idealized “hot/dry” boundary
limit; Tc (K) is the estimated Ts at the idealized “cold/wet” boundary;
and dT (K) is a pre-defined temperature difference between Th and Tc
for each pixel; it is a spatially dynamic constant that is unique for each
pixel and for day-of-year, calculated under clear-sky conditions for a
bare, dry soil where ET is assumed to be zero and sensible heat is as-
sumed to be maximum (Senay et al. 2013; Senay et al. 2016; Singh et
al. 2014).
The cold/wet limiting condition, Tc, is calculated from Ta based on the
assumption that for a well-watered vegetated pixel, Ta is equal to Tc, as
all the available energy is used for latent heat flux. However, a correc-
tion coefficient (c factor) is necessary for using Ta as a surrogate for Tc
for two reasons: 1) Ta and Ts will not correspond uniformly in magni-
tude even at the cold pixel due to differences in acquisition and process-
ing of the two datasets, and 2) daily maximum Ta from weather datasets
(TopoWx) and Ts from Landsat are acquired at different times of the day
(Senay et al. 2013). This difference can be rectified by evaluating the
ratio between Ta and Landsat Ts in the well-watered vegetated areas of
a study region where the assumption of Ta equal to Tc applies. As a result,
we determined the Tc boundary as a fraction of Ta using the c factor, in
Eq. (3):
Tc ¼ c ∗Ta ð3Þ
where c is the c factor that relates Ta to Ts on a well-watered, fully tran-
spiring vegetation surface. As described in the next section, we used a
unique c factor for each processed Landsat image. We calculated Th
(K) by adding dT (K) to Tc (K) using Eq. (4):
T h ¼ T c þ dT ð4Þ
Note that although dT is unique for each period and location, the
value does not vary from year to year (Senay et al. 2013). The observed
dT during the peak crop growing season is on the order of 20 K but
varies between 15 K and 25 K based on location (Senay et al. 2013).
For example, dT ranges from 13 K in late September to a maximum of
23 K in mid-July for the peak growing season of May–September in
PVID.
2.7.2. Calculation of the c factor
In previous studies (Senay et al. 2013; Senay et al. 2016; Singh et al.
2014), the median c factor was used to effectively characterize spatial
variation of ET with an acceptable level of seasonal stability. However,
historical analysis requires more accurate parameterization than cap-
turing seasonality or spatial pattern mapping. Particularly, initial evalu-
ation of ET over Palo Verde (Path/Row 38/37) produced saturated ETf,
i.e., well above 1.0 especially during the summer growing season due
to the use of a constant c factor. This necessitated the development of
a unique c factor for every image.
We calculated the c factor dynamically on a per-image basis using
the same general principle as reported in Senay et al. (2013) using Eq.
(5):
T S cold
c¼ ð5Þ
Ta
where Ts_cold (K) is the satellite-based Ts at the well-watered vegetated
pixels with model parameters as outlined below; Ta (K) is the corre-
sponding daily maximum air temperature.
102 G.B. Senay et al. / Remote Sensing of Environment 202 (2017) 98–112
We automated the process of determining a statistical c factor for
each image by implementing a set of constraining conditions to define
Ts_cold in Eq. (5). This creates an initial c factor called Tcorr (Ta correc-
tion), i.e., spatially explicit ratio of Ts to Ta that satisfy the constraints
listed in Table 2. Tcorr is the ratio of Ts to Ta and Tdiff is the difference be-
tween Ta and Ts. The Tcorr is then filtered by the following conditions: 1)
NDVI greater or equal to 0.7 and less than or equal to 1, 2) Ts N 270 K
(b270 K tend to be clouds or snow), 3) Tdiff between 0 and 30 K (in
most cases Tdiff stays between 3 and 10 K, but can be higher in the sum-
mer months, therefore, a bigger window is used to include as many po-
tential pixels as possible), and 4) at least 50 cloud-free pixels (as
determined by Fmask) that meet the above parameter requirements
are proposed for a statistically valid sample size (most cloud-free im-
ages will have several hundred thousand Landsat pixels, especially dur-
ing the growing season).
Exploratory analysis showed cloud-free images to have normal or
near-normal statistical distributions of Tcorr (not presented here); thus,
we determined the c factor of the Tcorr distribution by subtracting 2 stan-
dard deviations from the mean value using Eq. (6):
c ¼ T corr mean −2SD ð6Þ
where Tcorr_mean is the spatial mean of Tcorr in the image and 2SD is two
standard deviations of Tcorr pixels which satisfy the parameters in Table
2 in order to calibrate the c factor for the wettest and greenest locations.
Whenever an image failed to meet the c factor parameter constraints, a
historical scene-average was used based on all successful c factors from
31 years from that path/row. Additional discussion about c factor results
is presented in Section 3.1.
2.7.3. Cloud masking, gap filling, and monthly aggregation
The following is a brief overview on the procedure of masking
clouds, filling those masked pixels and aggregating the overpass ET
into cumulative totals, but full details can be found in Senay et al.
(2016). For this study, we used the Fmask algorithm to detect clouds
and cloud shadows for all Landsat imagery (Zhu et al. 2015). Identifying
cloud-contaminated pixels is essential as those masked out cloudy
pixels can then be interpolated with valid pixels. Pixel-based gap-filling
uses iterative linear interpolation based on cloud-free pixels from tem-
porally adjacent images in the path/row. When a given pixel ETf could
not be calculated due to clouds (or Landsat 7 data gaps), the interpolat-
ed or extrapolated ETf value for that pixel was used from previous or
next occurring image date(s) to fill the data gap. The extent of gap-fill-
ing exercise depends on the number of overpass scenes (within a
month, season, or year) and the number of days between consecutive
images. This resulted in a gap-filled ETf product that was spatially and
temporally consistent.
We created monthly and annual ET estimates from available ETf
values for each satellite overpass date. Aggregated ETa values were de-
rived from the scaled daily ETo and its nearest respective overpass ETf
value (Eq. 1) (Senay et al. 2016; Singh et al. 2011). This was a
Table 2
Constraining parameters for producing c factor by filtering the Tcorr. Pixels that meet all
listed criteria are used for c factor calculation.
Data Parameters Remark
NDVI 0.7 ≤ NDVI ≤ Healthy “greenest” vegetation
1.0
Ts N270 K Mask out clouds and snow pixels
Fmask ≥50 clear minimum statistical sample size
pixels
0 ≤ (Ta − Ts) Unrealistic Ta − Ts differences removed
Tdiff
≤ 30 K
Tcorr Ts/Ta Pixel-based preliminary c factor over well-vegetated
surfaces, meeting above criteria
c Mean Tcorr − c factor based on greenest and wettest/coldest vegetation
2SD
particularly useful method when using multiple satellites for a single
year, up to 45 images when combining Landsat 5 and 7 or combining
Landsat 7 and 8. In general, reasonable ET results can be determined
with at least 10–12 images per year and, ideally, at least one Landsat
image per month, especially during the crop growing season (Singh et
al. 2014). The early years of the study (1984 and 1985) had a limited
availability of cloud-free Landsat 5 images, numbering 7 images in
1984 and 10 images in 1985 with larger gaps between images. Howev-
er, for the entire historical study, the median number is 22 images per
year for each path/row. Most years have good representation in the
number of summer images (June–August) and have larger gaps (due
to cloudy images) in the winter months (November–January).
2.8. Basin-scale validation of ET estimates
We performed basin-scale validation of SSEBop ET estimates using
MPI ET data for all land cover types. In this study, the monthly MPI ET
data (mm/month) were summarized for eight HUC8 sub-basins and
PVID (Fig. 1). Monthly MPI ET time-series estimates were compared
with SSEBop ET data for 1984–2011. Coefficient of determination (R2)
and root mean square error (RMSE) were calculated for each study
site. The comparison over PVID was challenging since the PVID was rep-
resented by only three MPI pixels at 50 km resolution compared to sev-
eral thousand pixels for SSEBop ET at 100 m Landsat scale. To improve
the spatial correspondence we expanded the comparison extent outside
of PVID by including ETa results over the 3 MPI pixels.
Furthermore, the correspondence between remote sensing ETa and
crop consumptive use, generated from an inflow-outflow water balance
calculation by USBR, was evaluated using basin average annual values
for estimating bias and correlation statistics. The spatial average of an-
nual PVID ETa was calculated based on the PVID water accounting
basin boundary (excluding Cibola Valley and surroundings) and 2011
NLCD-identified cropland area within the California boundary
(90,606 ac); the USBR consumptive use estimate was obtained from
the Water Accounting Report on PVID water allocation and use
(http://www.usbr.gov/lc/region/g4000/wtracct.html, last accessed 2/
14/2017).
2.9. Year-to-year variability and seasonal Mann-Kendal trend analysis
One of the advantages of using Landsat is its relatively long record. In
addition to 31 years of monthly SSEBop ETa estimates from Landsat, we
also analyzed accompanying agro-climatic variables. Time-series infor-
mation (1984–2014) of spatially averaged monthly estimates of ETa,
ETo, precipitation (P), runoff (Q), NDVI, Ts, and Ta were derived for
each HUC8 sub-basin.
Year-to-year variability of the seven agro-climatic parameters was
quantified at monthly and annual time scales by using their respective co-
efficient of variability (CV), calculated as the ratio between standard devi-
ation and mean from 31 years of datasets for each month and annual
total. Furthermore, presence or absence of trend in time-series data for
each parameter was tested using the Seasonal Mann-Kendal (SMK)
trend test. The SMK trend test has been widely used for the analysis of
trends in hydrologic data (Douglas et al. 2000; Hirsch and Slack 1984;
Lettenmaier et al. 1994). This is a non-parametric rank-based method
for evaluating trends in time-series data. Hirsh and Slack (1984) defined
the SMK trend test using a set of equations as shown below.
2 3
X 11 X 12 … … X 1p
6 X 21 X 22 … … X 2p 7
6 7
X¼6
6 … … … … … 7 7 ð7Þ
4 …
… … … … 5
X n1 X n2 … … X np
where matrix X denotes a sequence of observations taken over p seasons
for n years. In this study, p = 12 months and n = 31 years. The SMK test
 G.B. Senay et al. / Remote Sensing of Environment 202 (2017) 98–112 103

examines slopes between all pairwise combinations of samples. The data
are first ranked with reference to time, and each data point is treated as
the reference for the data points in successive time periods. The SMK
test statistic for each month or a subset (Sg) is estimated as
n  
Sg ¼ ∑ sign X j −X i
ib j
101.5 gal (385 l per person per day) obtained from the USGS 2005
Water Use statistics (https://pubs.usgs.gov/circ/1344/pdf/c1344.pdf,
last accessed: 02/14/2017) and estimated 2.6 dwellers per household,
obtained from U.S. Census Bureau (https://www.census.gov/hhes/
families/files/graphics/HH-6.pdf, last accessed: 02/14/2017).
ð8Þ
3. Results and discussion

and 3.1. c factor

8
< 1; when : X j −X i N0 For the 3396 Landsat images processed, the success rate of the c fac-
signðX Þ ¼ 0; when : X j −X i ¼ 0 ð9Þ tor calculation (i.e. conditions met in Table 2) was 98% (3334). Remain-
:
−1; when : X j −X i b0 ing images featuring insufficient qualifying pixels were processed using
historical scene-based average c factors (as described at the end of
Xi and Xj are the observations at times i and j, respectively (i is the month Section 2.7.2) in order to include all usable pixels (albeit limited) even
and j is the same month from the following year) and n is total number of in cloudy images.
years. The overall SMK statistic (Sk) is estimated as sum of Sg shown as The use of this dynamic c factor approach across a large time series
g¼12
(31 years) has revealed important seasonal patterns and valuable in-
Sk ¼ ∑ Sg ð10Þ sights in SSEBop parameterization and its versatility for large-scale ap-
g¼1 plications. The two path/rows covering the PVID (path/row 38/37, 39/
37) exhibited the same seasonal pattern: around day of year (DOY)
where, g is the month/subset, Sk is asymptotically normal with a mean of 161 (10th June), the c factors “dip” and remain below a 0.96 average
0, and variance (var) is accounted for ties and serial dependency as until rising again after DOY 269 (26th September) as shown in Fig. 2.
We did not see the same seasonal pattern in the c factor generated
varðSk Þ ¼ ∑ σ g 2 þ ∑ σ gh ð11Þ from the other two path/rows covering parts of the Central Valley
g g; h;
g≠h (path/row 42/35, 43/34). Seasonally increased atmospheric water
vapor in this arid environment is one potential explanation for reducing
where g and h are subsets representing different months, σ2g =var[Sg] and the Ts signal (in relation to Ta) thereby lowering the c factor, however,
σgh =cov(Sg,Sh). more independent investigation is required to validate this hypothesis.
Finally, parameters such as SMK statistic and 5% significance level These varying differences in c factor highlight the importance of incor-
(p-value ≤ 0.05) presence or absence of trend were reported. Annual es- porating an image-based approach to take into account the spatio-tem-
timates for each parameter (1984–2014) were computed, and the slope poral variability of the c factor.
of the regression-fit obtained from each parameter was used to estimate
decadal rate of change (DROC). 3.2. Evaluation and validation of ET estimates

2.10. Fallowing and water saving economics 3.2.1. Basin-scale validation of ET estimates
Basin-scale validation of SSEBop ET data using MPI ET data for all
The linkages among water management decisions, hydrologic im- land cover types for HUC8 sub-basins is presented in Fig. 3 below. Re-
pacts, socio-economic transactions and implications were investigated sults indicate a reasonably high level of correlation with R2 ranging
by converting ET depth into volumetric magnitudes and applying pay-
out rates on a per acre foot (ac ft) basis and knowledge of household
water use rates.
We converted time-series ETa estimates (1984–2014) into irrigation
district-based volumetric water use. For this, we used the spatially ex-
plicit crop acreage (90,606 ac) map for the PVID (covering farm lands
within California) obtained from the NLCD 2011 dataset. This acreage
number is slightly (1.5%) smaller than the maximum acreage identified
by the USBR (91,990 ac) in 2000. To compute the historical mean water
use estimate, we excluded the years of the fallowing program (1992–
1994 and 2005–2014). The volumetric difference between the historical
mean of non-fallow years and a given year was used to calculate water
savings and district-wide payout amounts as a result of the fallowing
program. We used the 2014 cost of water ($752 per acre) obtained
from the PVID “landowner agreement” document (http://www.pvid.
org/MWDDocs/Landowner%20Agreement%2008-20-04%20rh.pdf, see
exhibit G, last accessed 01/04/2017) to estimate monetary equivalent
of water savings. Payout rates designed on a per acre basis were con-
verted into acre foot rates by using the long-term PVID mean ET depth
of 3.9 ft. (non-fallow years). Thus, an annual payout rate of $191/ac ft
of water saved was used as a basis to quantify the total potential pay-
ment to PVID farmers for not irrigating a portion of their land. The
value of savings for the earlier years would appear inflated as we used
a higher 2014 payout rate (inflation adjusted at 2.6%) for all years. How-
ever, a uniform payout rate would make this illustration comparable Fig. 2. Seasonality of c factor: (a) 872 historical (1984–2014) c factors processed from P/R:
across years using 2014 dollar value. The number of beneficiary house- 38/37 (PVID) and (b) 799 historical (1984–2014) c factors processed from P/R: 43/34
holds was calculated using an average per capita daily water use rate of (Central Valley) plotted over the course of a year.
104 G.B. Senay et al. / Remote Sensing of Environment 202 (2017) 98–112

Fig. 3. MPI Monthly ET compared to SSEBop monthly ET over 8 sub-basins within the Central Valley and the Palo Verde Irrigation District. PVID (bottom panel) also shows additional
SSEBop ET (grey line) that was made coarser to match the scale of MPI. For PVID, blue and grey data points (coarse scale) were used to calculate R2 and RMSE.

from 0.59 to 0.88 and an RMSE ranging from 14 to 24 mm/month for
Central Valley sub-basins. Fig. 3 indicates that SSEBop ET data (red-
dish-brown line) successfully captured the monthly dynamics and sea-
sonality seen in MPI data (blue line) in the Central Valley. Moreover, due
to the larger pixel size of MPI data, peak estimates of ET show more
smoothed features compared to the SSEBop ET.
The degree of correlation, however, appears to be dependent on the
spatial correspondence between the coarser resolution MPI data and
the more detailed landscape response represented by the SSEBop
Landsat data. For example, correlation was very low for the PVID. This
is a result of the spatial mismatch between the two datasets. As men-
tioned earlier, PVID is covered by three large pixels in MPI ET data com-
pared with several thousand pixels at Landsat scale from SSEBop ET
data. Fig. 3 (bottom panel) indicates the average estimate from three
pixels of MPI ET (blue line) and thousands of pixels of SSEBop ET (red-
dish-brown line). The larger pixel size of the MPI ET data included
large areas of near-zero ET estimates from the desert landscape outside
the PVID, decreasing the spatially averaged value for MPI ET. However,
to enable comparison we also plotted the mean estimate from SSEBop
ET (grey line) corresponding to the three coarse-pixels of MPI ET cover-
ing the PVID. Results indicate a low correlation (R2 of 0.11) and RMSE of
23.8 mm/year for PVID (blue versus grey line). The apparent differences
between the MPI ET and SSEBop ET could be attributed to the uncer-
tainties in the modeled SSEBop ET data (Velpuri et al. 2013) and the in-
accuracies in MPI ET to accurately quantify ET over irrigated areas due to
inadequate training data input into the upscaling model (Jung et al.
2009). Nevertheless, we have good confidence in our model result as
the magnitude of annual ET in the PVID is similar to that reported by
other researchers (Nagler et al. 2013). To further validate modeled ET
estimates over PVID, we compared SSEBop ET with water balance and
crop acreage estimates obtained from the USBR.
3.2.2. PVID ET and USBR water balance ET
The year to year correspondence between SSEBop ET and USBR con-
sumptive use is encouraging in capturing major annual patterns with a
significant positive correlation (r = 0.71, p b 0.01), but does not track in
the same direction in some years (Fig. 4). Notable differences are when
SSEBop ET is either higher in consecutive years 1990–1993 or lower
 G.B. Senay et al. / Remote Sensing of Environment 202 (2017) 98–112
Fig. 4. Historical (1984–2014) traces of annual ET over PVID using two methods: 1) SSEBop ET is satellite-based ET and 2) water balance ET is estimated from the differences between
inflows and outflows, which are reported as consumptive use by USBR. Crop acreage is also shown with period of fallowing initiated under the PVID and Metropolitan Water District
of Southern California program. Scatterplot shows relationships between the USBR water balance consumptive use and SSEBop ET. Precipitation is annual total from PRISM.
during 1999–2002. However, the average SSEBop ET (1131 mm) and
USBR ET (1070 mm) show a close match within 6% of each other, sug-
gesting the reliability of the estimated volumetric crop water use
estimate.
Annual ET statistics for the PVID show a general decline in ET over
the 1984–2014 period, peaking at 1335 mm in 1995 but falling to a
low of 840 mm in 2011. The declining trend is attributable to land man-
agement practices rather than climatic forces as precipitation has mini-
mal impact in the PVID, averaging just over 92 mm a year and never
reaching higher than 209 mm. The association between crop acreage
and ET is demonstrated by a strong correlation coefficient (r) of 0.84
and 0.87 with SSEBop ET and USBR ET, respectively. There is a sharp
and sustained decline in annual ET from 2005 through 2011 in associa-
tion with a decline in farmed acreage reported by the PVID
Watermaster, as well as a decline in 1993–1994 that corresponds to a
Fallow Field program implemented by MWD/PVID (http://pvid.org/
PVID-MWD.html, accessed 02/14/2017) — see Fig. 4. The consumptive
water use values (water balance) reported by the USBR for the PVID
are mainly a result of the difference between water diverted into the
PVID (inflow + precipitation) and loss and return flow (outflow). How-
ever, not all of this can be attributed to crop water use as there are a few
small towns in the PVID such as Blythe and Ripley, CA. But, the PVID is
predominantly agricultural, where farmlands account for about N72%
of the PVID boundary (within California). In addition, the USBR water
balance ET values largely track well with the annual SSEBop ET patterns
over NLCD-identified croplands, especially during the MWD/PVID Fal-
low Field program with a substantial dip in both annual ETa and USBR
water balance ET since 2005 (start of the fallowing program) and rising
together in later years (2012–2014) with the easing of the of fallowing
program (Fig. 4). This finding suggests that the historical ET patterns of
irrigated agricultural lands of PVID can represent the response to land
management practices and their impact on water use.
3.3. Spatio-temporal dynamics of ETa
This study used, for the first time, 31 years of Landsat imagery to un-
derstand the dynamics of water use in the PVID and Middle and Lower
Central Valley sub-basins. Fig. 5 shows the median and annual ETa esti-
mates for the PVID computed from 31 years of analysis. The map shows
a range of ETa with higher values coming from agricultural fields and
along riparian corridors and lower estimates coming from non-agricul-
tural areas. The time-series information on ETa (shown for every
5 years) clearly reflects the impact of the fallowing program (1993–
1994 and since 2005) on the reduction of crop water use as shown by
105
the reduced ET (absence of irrigation on several fields) in 2005 and
2010 compared to earlier years (Figs. 4 and 5).
Fig. 6 shows the median and annual ETa estimates for the Middle and
Lower Central Valley HUC8 sub-basins computed from 31 years of data.
It is to be noted that unlike the PVID, these sub-basins cover a variety
of landscapes, both agricultural and non-agricultural. The annual ETa
in these hydrologic sub-basins points to distinct regions: high ETa in
the cropped valley centers and forested headwaters separated by
low ETa in the drier foothills located in the east. The overall ETa mag-
nitude was found to be less than PVID estimates, a result of a relative-
ly lower atmospheric demand (ETo). The time-series information on
ET a (shown for every 5 years) is also illustrated for the Lower San
Joaquin sub-basin. The zoom-in views cover mostly agriculture
areas where greater spatio-temporal dynamics of ET a can be seen
(Fig. 6).
3.3.1. Characterizing temporal and spatial variability of agro-climatic/hy-
drologic variables
Inter-comparison and characterization of time-series variability and
trends can provide useful information for effective management of
water. Fig. 7 shows the comparison of annual time-series data for the
climatic (ETo, P, and Ta) and management (ETa, Q, NDVI, and Ts) influ-
enced agro-hydrologic parameters across the study sites. Climate in
the PVID is more arid than in the Central Valley sub-basins, and the at-
mospheric evaporative demand as measured by ETo for the PVID was
found to be higher (~ 2200 mm/year, Coefficient of Variation, CV =
3%) than Central Valley sub-basins (~1550 mm/year, CV = ~4.5%), indi-
cating that irrigation in the PVID would require more water for the same
crop. Mean annual precipitation decreases from a high of 310 mm in the
Middle Central Valley to a low of 93 in the PVID, with 199 mm for the
Lower Central Valley. On the other hand the year-to-year variability as
measured by CV increases from a low of 35% for the Middle Central Val-
ley to a high of 50% in the PVID, with 40% for the Lower Central Valley.
The PVID showed the highest annual average for daily maximum air
temperature (Ta of 305 K, CV = 0.19%) among all study sites. Lower
and Middle Central Valley sub-basins showed average annual Ta of
299 K (CV = 0.21%) and 298 K (CV = 0.22%), respectively.
The average annual ETa for the PVID (including regions from Califor-
nia and Arizona) was found to be 1208 mm with a higher year-to-year
variability (CV = 11%) than the Central Valley sub-basins, with mean
annual ET and CV of 705 mm and 9%, and 772 mm and 6% for the Middle
and Lower sub-basins, respectively. The ETa to ETo ratio for the PVID,
Lower, and Middle Central Valley sub-basins was found to be 55%,
46%, and 49%, respectively. Runoff (not shown) in the PVID was found
to be low (6 mm/year, CV = 140%) with an average annual runoff
106 G.B. Senay et al. / Remote Sensing of Environment 202 (2017) 98–112

Fig. 5. Median annual ETa estimates (SSEBop) for the Palo Verde Irrigation District over 1984–2014 (left panel). Also shown are maps of annual ETa for every 5 years since 1985.

coefficient (RC: ratio of annual runoff to precipitation) of 7%. On the
other hand, Lower (56 mm, CV = 96%) and Middle (81 mm, CV =
91%) Central Valley sub-basins showed relatively higher estimates of
runoff, with a substantial year-to-year variability (CV close to 100%)
but still lower than the PVID. The average annual RC for Lower and Mid-
dle Central Valley sub-basins was found to be much higher than the
PVID with RCs of 27% and 25%, respectively. Ts showed a difference of
up to 5 K among sites with annual Ts estimates of 303 K (CV = 0.75%),
300 K (CV = 0.83%), and 298 K (CV = 0.73%) for the PVID, Lower, and
Middle Central Valley sub-basins, respectively. Average annual NDVI
for the study sites was 0.38, 0.35, and 0.36, respectively. Temporal pat-
terns of NDVI in the PVID closely followed that of ETa where the majority
of the sub-basin is covered by irrigated crops.
Using differences in coefficient of variation (CV), we can conclude
that management influenced agro-hydrologic variables (ETa, Q, and Ts)
tend to have a much higher (up to 4 times) year-to-year variability
than their corresponding climatic variables (ETo, P, and Ta). This is not
necessarily attributed to the sole impact of management, but reveals
differences in process characteristics between climatic drivers and
agro-hydrologic response variables. For example, air temperature (cli-
mate) is a regional variable and tends to be stable over the years as com-
pared to land surface temperature which is a local variable that can be
influenced by moisture movement and vegetation conditions caused
by either or both natural- and human-influenced processes. However,
management can substantially influence agro-hydrologic variables
with little or no impact to climatic drivers at such local scales. This is

Fig. 6. Median annual ETa estimates (SSEBop) for the Middle and Lower Central Valley sub-basins over 1984–2014. Also shown are zoom-in views of time-series annual ETa from the
irrigated areas located in the Lower San Joaquin sub-basin for every 5 years since 1985.
 G.B. Senay et al. / Remote Sensing of Environment 202 (2017) 98–112 107

Fig. 7. Inter-comparison of annual time-series observations (1984–2014) for all parameters studied for the Palo Verde Irrigation District, Middle Central Valley sub-basins (Rock Creek,
Lower San Joaquin, Middle San Joaquin and, Fresno River) and Lower Central Valley sub-basins (Upper Kaweah, Upper Tule, Upper Deer and, Upper Poso). Note: For estimates of slope
and significance of trends, please refer to Table 4 on decadal rate of change. μ and CV are the mean and coefficient of variation, respectively.

more evident in the PVID when a drastic management change such as a
fallowing program caused a significant decline in crop water use there-
by exaggerating the annual variability in ETa compared to ETo.
3.3.2. Characterizing seasonality of agro-climatic/hydrologic variables
within agricultural lands
Characterizing annual time-series data provides information on
long-term trends, but does not provide information on the seasonal
(intra-annual timing) variability of water use drivers and responses.
Fig. 8 compares seasonality of the agro-climatic variables using mean
monthly values in the PVID (entire PVID boundary including regions
of California and Arizona), Lower, and Middle Central Valley sub-basins.
The climatic parameter ETo shows similar seasonality across all study
sites including low estimates in winter months (November–January)
and high estimates in summer months (June–August). Average ETo esti-
mates in summer were found to be 285 mm, 226 mm, and 234 mm/-
month for the PVID, Lower, and Middle Central Valley sub-basins,
respectively. High month-to-month precipitation variability was ob-
served in all study sites, with mean monthly precipitation ranging
from b0.1 mm in June for the PVID to a high of 59 mm in January for
Middle Central Valley sub-basins. Although the peak precipitation oc-
curred in January for all three study sites, 31% of the annual precipitation
in the PVID is contributed by the summer months of July–August–Sep-
tember as opposed to only 1% for the Central Valley sub-basins. As ex-
pected, Ta showed the highest temperatures in summer months and
lowest temperatures in winter months. Like in the annual time scale,
108 G.B. Senay et al. / Remote Sensing of Environment 202 (2017) 98–112

Fig. 8. Mean monthly seasonality of agro-climatic variables for the Palo Verde Irrigation District, Middle Central Valley sub-basins, and Lower Central Valley sub-basins. Dotted line
indicates the mean and shaded grey area shows the variability (±1SD) observed during 1984–2014.

historical monthly Ta variability was found to be lowest when compared
to other agro-climatic variables (shaded area in Fig. 8).
ETa followed a similar pattern as ETo, with peak estimates in summer
months and low estimates in winter months. However, magnitude of
basin average ETa was higher in the PVID, followed by Middle and
Lower Central Valley sub-basins with corresponding peak-month
values of 211 mm, 144 mm, and 136 mm, respectively. Monthly runoff
is negligible (b1 mm) in the case of the PVID. However, Lower and Mid-
dle Central Valley sub-basins showed high month-to-month variability
with the peak runoff occurring in April with 8.5 mm and 12.1 mm for
the Lower and Middle Central Valley, respectively; the lowest runoff oc-
curred in November with b3 mm per month. The historical spread for
each month is more pronounced for runoff compared to the precipita-
tion for Central Valley sub-basins, especially for the rainy winter
months, but the monthly runoff spread for the PVID is not as clear due
to smaller magnitudes, exhibiting an overall comparable CV between
precipitation and runoff at the monthly time scale. Ts also showed sim-
ilar seasonal pattern as Ta but with slightly more spread for each month.
Monthly NDVI plots provided information on crop phenology in
each study site. Although average NDVI for the PVID and Lower and
 G.B. Senay et al. / Remote Sensing of Environment 202 (2017) 98–112 109

Middle Central Valley sub-basins is comparable, they have distinct dif-
ferences in month-to-month phenological curves. The PVID showed
more of a unimodal curve implying a dominant single season cropping
pattern. Lower and Middle Central Valley sub-basins showed distinct bi-
modal curves implying more double cropping. It is interesting to note
that despite the bimodal crop phenological curves, ETa shows only a
unimodal signal, indicating the relatively small contribution of winter
ET to the overall annual water use. Comparing the monthly relative
spread (as measured by CV) with the annual spread for each of the cli-
mate- and management-influenced agro-hydrologic variables, the agro-
hydrologic variables still maintained a higher degree of variability. This
observation is not as clear for the PVID runoff where monthly precipita-
tion shows as many months with higher CV in precipitation than runoff.
This also implies the impact of management on the precipitation-runoff
process is relatively small when compared to its impact on land surface
temperature and ET, especially in the PVID where the peak runoff in
March constitutes only a small percentage (1.1%) of the corresponding
monthly ET and markedly lower during the summer.
3.4. SMK trend analysis
Results from the SMK test conducted on monthly time-series data
for agricultural lands are presented in Table 3. PVID climatic parameters,
such as ETo and Ta, showed a significant positive trend during 1984–
2014. However, precipitation did not show a similar trend. The effect
of management practices can be understood based on trends observed
in management-influenced agro-hydrologic parameters. ETa exhibited
a significant decreasing trend. Although an increase in ETo was found,
management practices such as fallowing programs may have decreased
basin-wide ET losses. However small in absolute magnitude, runoff co-
efficient decreased from N 20% at the beginning of the study period to
Table 3
Seasonal Mann-Kendall trend analysis (1984–2014) of basin average agro-climatic/hydro-
logic parameters summarized for nine study sites.
b5% of precipitation towards the end. The negative trend in runoff, de-
spite a neutral precipitation, may suggest a reduction in diversion and
return flow from the irrigated fields as supported by the lower ETa. In-
terestingly, NDVI did not show any significant trend, while Ts showed
a positive trend as expected with the reduction in ETa.
Middle Central Valley sub-basins were similar to one another, show-
ing no significant trend in ETo or precipitation but with a significant pos-
itive trend in Ta. On the other hand, ETa exhibited a decreasing trend in
Rock Creek and an increasing trend in Fresno River sub-basin. Contrary
to what is expected from the increasing ETa, the total area under irriga-
tion has reduced in Fresno River sub-basin since 1989, according to the
Fresno Irrigation District Water Management Plan (http://www.usbr.
gov/mp/watershare/wcplans/2015/fresno-irrigation-district.pdf, last
accessed 7/26/2016). However, the contrary increasing trend in basin
ETa may indicate a potential switch to water-intensive crops. Further-
more, we did not see any significant trend in runoff. NDVI showed an in-
crease in Rock Creek while it did not show any trend in the other three
Middle Central Valley sub-basins.
Lower Central Valley sub-basins also showed similar trends in cli-
matic parameters with ETo and P having no trend while Ta showed a sig-
nificant positive trend. All sub-basins showed similar impacts of
management practices with no trend in ETa and Q, but positive trends
in NDVI and Ts (except for Upper Poso sub-basin which showed a posi-
tive trend in ETa). The contrary increase in NDVI accompanied with an
increasing Ts, or an increasing NDVI accompanied by a decreasing ETa
could not be explained with the available data.
Results show that in all nine sub-basins studied, temperature (Ta and
Ts) consistently showed a positive trend. These findings are similar to
other researchers (Bonfils et al. 2007; Westerling et al. 2006) who dem-
onstrated a warming trend in the California region. Lack of trend in pre-
cipitation in these sub-basins is also in line with findings from Velpuri
and Senay (2013), which indicated absence of precipitation trends in
the western U.S.
3.5. Decadal rate of change

Climate parameters Agro-hydrologic parameters

Sub-basin Decadal rate of change for each parameter estimated from the annu-
name Statistic ETo P Ta ETa Q NDVI Ts al estimates of agro-climatic variables is presented in Table 4. For cli-
PVID MKS 774 −63 798 −1020 −1846 −654 1456 matic variables, the PVID showed increasing ETo and Ta. ETo increased
p-Val 0.027 0.761 0.002 0.029 0.000 0.184 0.000 up to 26 mm/decade and a total of nearly 80 mm over the 31-year peri-
Trend (+) (#) (+) (−) (−) (#) (+) od (1984–2014). Ta increased at a rate of 0.4 K/decade and a total of
Middle Central Valley sub-basins 1.2 K over the 31-year period. While P showed a decline of up to 11
Rock Creek MKS 612 −74 670 −616 −384 −462 1392 mm/decade, it was not statistically significant. On the other hand, a gen-
p-Val 0.069 0.696 0.013 0.038 0.542 0.116 0.000 eral declining trend of parameters such as ETa, Q, and NDVI reflected im-
Trend (#) (#) (+) (−) (#) (#) (+)
Lower San MKS 620 48 652 −124 −411 641 1408
pacts of management practices in the PVID. A decline in ETa up to 58
Joaquin p-Val 0.075 0.819 0.020 0.684 0.514 0.019 0.000 mm/decade or 180 mm over 31-year period indicates changes in irriga-
Trend (#) (#) (+) (#) (#) (+) (+) tion and land management practices. This decline in ETa (5.8 mm/year)
Middle San MKS 548 78 610 −452 −513 815 1458 is equal to 2584 acre-feet (3.19 M m3) of water savings from the entire
Joaquin p-Val 0.125 0.714 0.025 0.202 0.406 0.002 0.000
Trend (#) (#) (+) (#) (#) (+) (+)
Table 4
Fresno River MKS 507 84 632 766 −581 1443 1001
Decadal rate of change for agro-climatic/hydrologic parameters summarized by study site.
p-Val 0.150 0.682 0.020 0.018 0.342 0.000 0.000
Trend (#) (#) (+) (+) (#) (+) (+) Climate parameters Agro-hydrologic parameters
Lower Central Valley sub-basins Sub-basin ETo P Ta ETa Q NDVI Ts
Upper MKS 310 −44 666 −396 −651 1213 1198 Units (per decade) mm mm K mm mm (−) K
Kaweah p-Val 0.311 0.809 0.015 0.273 0.303 0.000 0.000
Trend (#) (#) (+) (#) (#) (+) (+) PVID 26⁎ −11 0.4⁎ −58⁎ −5.2⁎ −0.02⁎ 0.8⁎
Upper Tule MKS 186 −258 576 −394 −625 955 910 Middle Central Valley sub-basins
p-Val 0.517 0.219 0.025 0.272 0.324 0.000 0.003 Rock Creek 25 −7.3 0.3⁎ −24⁎ −6.3 −0.01⁎ 0.4
Trend (#) (#) (+) (#) (#) (+) (+) Lower San Joaquin 23 8.4 0.3⁎ 0.11 −6.3 −0.01 0.3
Upper Deer MKS 232 −11 674 258 −610 1163 1048 Middle San Joaquin 20 0.98 0.3⁎ −16 −2.4 0.02⁎ 0.5
p-Val 0.442 0.961 0.013 0.505 0.335 0.001 0.001 Fresno River 16 −2.1 0.3⁎ 35⁎ −2.3 0.00 −0.2
Trend (#) (#) (+) (#) (#) (+) (+)
Upper Poso MKS 402 −258 576 952 −625 925 866 Lower Central Valley sub-basins
p-Val 0.190 0.219 0.025 0.016 0.324 0.000 0.004 Upper Kaweah 6.9 −4.8 0.3⁎ −15 −5 0.00 0.2
Trend (#) (#) (+) (+) (#) (+) (+) Upper Tule 2.8 −3.9 0.3⁎ −14 −4.9 0.01 0.2
Upper Deer 7.1 −1.9 0.3⁎ −1.7 −4.9 0.00 0.2
Note: MKS: Mann-Kendall Statistic; p-Val: statistical p-value; (#): no significant trend;
Upper Poso 13 −6.2 0.3⁎ 35⁎ −4.9 0.00 −0.2
(+): significant positive trend; (−): significant negative trend. Significance determined
at p = 0.05. ⁎ Indicates statistical significance at p = 0.05.
110 G.B. Senay et al. / Remote Sensing of Environment 202 (2017) 98–112
PVID — see Fig. 9. This reduction could mainly be attributed to the
fallowing program in the irrigation district (see Fig. 4).
Middle Central Valley sub-basins all showed no statistically signifi-
cant trend in ETo and P. However, Ta increased significantly (0.3 K/de-
cade). In terms of agro-hydrologic parameters for Rock Creek, ETa and
NDVI showed a negative trend with ETa declining at 24 mm/decade.
However, Q and Ts did not show any trend. In the Lower San Joaquin
sub-basin, none of the parameters (ETa, Q, NDVI, Ts) had any significant
decadal trend, indicating the absence of a significant change in manage-
ment practices. In the Middle San Joaquin sub-basin, only NDVI showed
significant increase, while other parameters did not show any signifi-
cant decadal change. In the Fresno River sub-basin, agro-hydrologic pa-
rameters (Q, NDVI, and Ts) did not show any significant decadal change.
However, ETa showed significant positive DROC with an increasing
trend of 35 mm/decade. This shows that the consumptive water use in-
creased by 14% in Fresno River sub-basin over 31 years.
Lower Central Valley sub-basins also showed similar climatic trends
with only Ta showing increasing decadal change (0.3 K/decade). ETo and
P did not show significant decadal change. The impacts of management
in these sub-basins were found to be minimal as all of the parameters
showed no significant trends, except Upper Poso, which exhibited an in-
creasing DROC of ETa at 35 mm/decade.
3.6. Estimation of water use statistics for agricultural lands in the PVID
One of the major advantages of using Landsat imagery for ETa esti-
mation is its ability to provide field-level information for district-wide
crop water use statistics. Here we illustrate the use of time-series
Landsat imagery to compute volumetric water use statistics for the
PVID. Fig. 9 (top panel) shows the long-term time series of water use
(in × 1000 ac ft). The historical mean water use in the PVID is shown
as a dashed line. According to SSEBop ET, the maximum and minimum
water use in the PVID occurred in 1995 (398,100 ac ft in a non-fallow
year) and 2011 (249,700 ac ft, fallow year), respectively. Fig. 9 (bottom
left axis) indicates volumetric water use deviations from the non-fallow
years. Positive bars indicate years when water is saved from a basin-
wide reduction in ET. The largest saving occurred in 2011 with an esti-
mated 107,200 ac ft. Fig. 9 (bottom panel, right axis 1) indicates mone-
tary value of water savings (in millions of U.S. Dollars); this is the annual
payout to PVID farmers for not using irrigation water at the rate of $191
per ac ft, with a maximum total payment of $20.5 million in 2011. The
Fig. 9. Time-series annual estimates in the PVID: (top) volumetric crop water use (ac ft) with mean of non-fallow years shown as dotted line; (bottom left) relative water savings due to
fallowing as deviations from mean of non-fallow years (×1000 of ac ft); and (bottom right 1) estimated monetary value of water saved (million $) and (bottom right 2) number of
beneficiary households (x 1000).
water saved from agricultural use is redirected for municipal water
use through the Metropolitan Water District of Southern California. As-
suming all water is used for domestic purposes, the estimated number
of beneficiary households is shown in Fig. 9 (bottom right axis 2),
reaching up to 362,700 households or nearly 1 million individuals in
2011. This illustrates the strong linkages between upstream water man-
agement decisions and broader hydrologic and societal impacts from
farming communities to municipal dwellers.
4. Conclusions
Our main objective was to quantify and characterize the spatio-tem-
poral dynamics of historical (1984–2014) ET using Landsat images over
the Palo Verde Irrigation District and irrigated sub-basins in parts of the
Central Valley in California along with six other agro-climatic/hydrolog-
ic variables.
A total of 3396 Landsat images were processed and analyzed to pro-
duce monthly and annual ET using the SSEBop model with a unique c
factor parameterization for each Landsat image. The c factor helps estab-
lish the lower limit of the land surface temperature (Tc) using a daily
maximum air temperature (Ta). Our results highlight the importance
of a dynamic c factor to account for the observed seasonal and locational
variation in the ratio of Ts (from well-watered, full canopy, healthy veg-
etation) at the time of image acquisition (morning for Landsat) to that of
Ta (daily maximum), which generally occurs in the afternoon.
A reasonable correspondence between SSEBop ET and USBR con-
sumptive use reports (water balance approach) and MPI ET (basin val-
idation) indicates the reliability of Landsat-based ET for trend analysis of
crop water use. In addition, since Landsat-based ET can be estimated at
the field scale, the relationships between water allocation and use can
be analyzed at the individual field level.
We grouped seven key agro-climatic/hydrologic parameters into
two: climate and agro-hydrologic groups. Climatic parameters included
atmospheric water demand (ETo), air temperature (Ta) and precipita-
tion (P), and agro-hydrologic variables consisted of actual ET (ETa),
land surface temperature (Ts), runoff (Q) and NDVI. Generally, agro-hy-
drologic parameters showed higher year-to-year variability than cli-
matic parameters, indicating the combined impact of dynamic local-
scale land surface processes and management practices on agro-hydro-
logic responses as compared to the more stable regional processes on
climatic drivers. In terms of long-term changes, only Ta, among climatic
 G.B. Senay et al. / Remote Sensing of Environment 202 (2017) 98–112
parameters, had a consistent positive trend (an increase of up to 1.2 K
over 31 years) in southern California. ETo and P showed a generally neu-
tral trend. On the other hand, among agro-hydrologic parameters, Ts
showed a consistent positive trend while NDVI indicated a positive
trend for the majority of sub-basins. However, no consistent trend
was observed for ETa and Q across different sites. But a negative ETa
trend despite a positive ETo trend in the PVID reveals the effect of land
and water management practices in reducing actual crop water use,
confirming the hydrologic impacts of the PVID/MWD fallowing pro-
gram. Some sub-basins showed unexpected, contrary trends such as
positive trends for both NDVI and Ts which require further investigation.
This study reinforces the usefulness of remote sensing-derived infor-
mation to monitor and assess the impact of management decisions on
water use and associated economic transactions. For example, during
the peak crop fallowing program, water savings reached a maximum
of nearly 107,200 ac ft in 2011, which is equivalent to a payout of nearly
$20.5 million for PVID farmers. This in turn can be translated into a po-
tential provision of adequate water supply to 362,700 households or ~1
million individuals in 2011.
Our study demonstrated the application of “Big Data” science in his-
torical Landsat data and how large-scale spatio-temporal analytics can
decipher patterns in crop water use dynamics in regards to linkages
among water management decisions, hydrologic processes and eco-
nomic transactions. This effort was made possible by using a combina-
tion of thousands of freely available Landsat images, gridded weather
datasets, and a simplified, robust ET modeling algorithm such as
SSEBop. Similar approaches can be adopted and implemented by
land and water managers in different parts of the world in a cost ef-
fective way in order to make meaningful and informed management
decisions.
Acknowledgements
This work was performed under the U.S. Geological Survey (USGS)
contracts G13PC00028 and G10PC00044 in support of the USGS Land
Change Science projects such as Landsat Water Balance and
WaterSMART. We gratefully acknowledge the institutions and individ-
uals who made various geospatial data freely available: Landsat (USGS
Earth Resources Observation and Science (EROS) Center); TopoWx air
temperature (North Central Climate Science Center through Dr. Jared
Oyler); GridMET reference evapotranspiration (University of Idaho
through Dr. John Abatzoglou); MPI fluxnet data (Dr. Martin Jung); U.S.
Bureau of Reclamation (USBR) Consumptive User report. We would like
to extend our appreciation to all peer reviewers (Dr. Trent Biggs, Dr.
Bruce Wylie and three other anonymous reviewers) whose constructive
and detailed comments improved the final manuscript. Any use of trade,
firm, or product names is for descriptive purposes only and does not im-
ply endorsement by the U.S. Government.
References
Abatzoglou, J.T., 2013. Development of gridded surface meteorological data for ecological
applications and modelling. Int. J. Climatol. 33:121–131. http://dx.doi.org/10.1002/
joc.3413.
Alley, W.M., Evenson, E.J., Barber, N.L., Bruce, B.W., Dennehy, K.F., Freeman, M.C., Freeman,
W.O., Fischer, J.M., Hughes, W.B., Kennen, J.G., Kiang, J.E., Maloney, K.O., Musgrove, M.,
Ralston, B., Tessler, S., Verdin, J.P., 2013. Progress Toward Establishing a National As-
sessment of Water Availability and Use. 1384 (p. 34).
Anderson, M.C., Kustas, W.P., Norman, J.M., Hain, C.R., Mecikalski, J.R., Schultz, L.,
González-Dugo, M.P., Cammalleri, C., d'Urso, G., Pimstein, A., Gao, F., 2011. Mapping
daily evapotranspiration at field to continental scales using geostationary and polar
orbiting satellite imagery. Hydrol. Earth Syst. Sci. 15:223–239. http://dx.doi.org/10.
5194/hess-15-223-2011.
Anderson, R.G., Lo, M.H., Famiglietti, J.S., 2012. Assessing surface water consumption using
remotely-sensed groundwater, evapotranspiration, and precipitation. Geophys. Res.
Lett. 39, L16401. http://dx.doi.org/10.1029/2012GL052400.
Bastiaanssen, W., Karimi, P., Rebelo, L.-M., Duan, Z., Senay, G., Muttuwatte, L., Smakhtin,
V., 2014. Earth observation based assessment of the water production and water con-
sumption of Nile Basin agro-ecosystems. Remote Sens. 6:10306–10334. http://dx.doi.
org/10.3390/rs61110306.
111
Bonfils, C., Duffy, P.B., Santer, B.D., Wigley, T.M.L., Lobell, D.B., Phillips, T.J., Doutriaux, C.,
2007. Identification of external influences on temperatures in California. Clim.
Chang. 87:43–55. http://dx.doi.org/10.1007/s10584-007-9374-9.
Chatterjee, S., Sritharan, S., Watts, D., Tarpey, T., 2012. Estimating Palo Verde irrigation
district evapotranspiration using MODIS enhanced vegetation index and micro-mete-
orological data. In: Neale, C.M.U., Cosh, M.H. (Eds.), IAHS-AISH Publication 352. Inter-
national Association of Hydrological Sciences, UK, pp. 406–409.
Daly, C., Taylor, G.H., Gibson, W.P., Parzybok, T.W., Johnson, G.L., Pasteris, P.A., 2000. High-
quality spatial climate data sets for the United States and beyond. Trans. ASAE 43:
1957–1962. http://dx.doi.org/10.13031/2013.3101.
Daly, C., Smith, J.W., Smith, J.I., McKane, R.B., 2007. High-resolution spatial modeling of
daily weather elements for a catchment in the Oregon Cascade Mountains, United
States. J. Appl. Meteorol. Climatol. 46:1565–1586. http://dx.doi.org/10.1175/
jam2548.1.
Douglas, E.M., Vogel, R.M., Kroll, C.N., 2000. Trends in foods and low flows in the United
States: impact of spatial correlation. J. Hydrol. 240, 90–105.
Elhaddad, A., Garcia, L.A., Chávez, J.L., 2011. Using a surface energy balance model to cal-
culate spatially distributed actual evapotranspiration. J. Irrig. Drain. Eng. 137:17–26.
http://dx.doi.org/10.1061/(ASCE)IR.1943-4774.0000276.
Felton, C.M.M., 1979. Water usage and microclimate. Water Air Soil Pollut. 12:91–101.
http://dx.doi.org/10.1007/BF01060528.
Fisher, J.B., Baldocchi, D.D., Misson, L., Dawson, T.E., Goldstein, A.H., 2007. What the
towers don't see at night: nocturnal sap flow in trees and shrubs at two AmeriFlux
sites in California. Tree Physiol. 27, 597–610.
U.S. Geological Survey, 2015. Landsat-Earth observation satellites. Fact Sheet:
pp. 2015–3081 http://dx.doi.org/10.3133/fs20153081.
Hirsch, R.M., Slack, J.R., 1984. A nonparametric trend test for seasonal data with serial de-
pendence. Water Resour. Res. 20, 727–732.
Homer, C., Dewitz, J., Yang, L., Jin, S., Danielson, P., Xian, G., Coulston, J., Herold, N.,
Wickham, J., Megown, K., 2015. Completion of the 2011 National Land Cover Data-
base for the conterminous United States — representing a decade of land cover
change information. Photogrammetric Engineering and Remote Sensing:
pp. 346–354 http://dx.doi.org/10.14358/PERS.81.5.345.
Jian, X., Wolock, D., Lins, H., 2008. WaterWatch—Maps, Graphs, and Tables of Current, Re-
cent, and Past Streamflow Conditions. 3031.
Jin, J., Miller, N.L., Yang, B., 2012. Modeling of the effects of land-use change and irrigation
on local climate. Land Use: Planning, Regulations, and Environment. Nova Science
Publishers, New York, pp. 125–147.
Jung, M., Reichstein, M., Bondeau, A., 2009. Towards global empirical upscaling of
FLUXNET eddy covariance observations: validation of a model tree ensemble ap-
proach using a biosphere model. Biogeosciences 6, 2001–2013.
Jung, M., Reichstein, M., Ciais, P., Seneviratne, S.I., Sheffield, J., Goulden, M.L., Bonan, G.,
Cescatti, A., Chen, J., de Jeu, R., Dolman, A.J., Eugster, W., Gerten, D., Gianelle, D.,
Gobron, N., Heinke, J., Kimball, J., Law, B.E., Montagnani, L., Mu, Q., Mueller, B.,
Oleson, K., Papale, D., Richardson, A.D., Roupsard, O., Running, S., Tomelleri, E.,
Viovy, N., Weber, U., Williams, C., Wood, E., Zaehle, S., Zhang, K., 2010. Recent decline
in the global land evapotranspiration trend due to limited moisture supply. Nature
467:951–954. http://dx.doi.org/10.1038/nature09396.
Lettenmaier, D.P., Wood, E.F., Wallis, J.R., 1994. Hydro-climatological trends in the conti-
nental United States, 1948–88. J. Clim. 7, 586–607.
Lo, M.H., Famiglietti, J.S., 2013. Irrigation in California's Central Valley strengthens the
southwestern U.S. water cycle. Geophys. Res. Lett. 40:301–306. http://dx.doi.org/10.
1002/grl.50108.
Lobell, D., Bala, G., Mirin, A., Phillips, T., Maxwell, R., Rotman, D., 2009. Regional differences
in the influence of irrigation on climate. J. Clim. 22:2248–2255. http://dx.doi.org/10.
1175/2008jcli2703.1.
Marshall, M., Thenkabail, P., Biggs, T., Post, K., 2016. Hyperspectral narrowband and mul-
tispectral broadband indices for remote sensing of crop evapotranspiration and its
components (transpiration and soil evaporation). Agric. For. Meteorol. 218-219:
122–134. http://dx.doi.org/10.1016/j.agrformet.2015.12.025.
Michelsen, A.M., Jones, S., Evenson, E., Blodgett, D., 2016. The USGS water availability and
use science program: needs, establishment, and goals of a water census. J. Am. Water
Res. Assoc. http://dx.doi.org/10.1111/1752-1688.12422.
Mueller, N., Lewis, A., Roberts, D., Ring, S., Melrose, R., Sixsmith, J., Lymburner, L.,
McIntyre, A., Tan, P., Curnow, S., Ip, A., 2016. Water observations from space: map-
ping surface water from 25 years of Landsat imagery across Australia. Remote Sens.
Environ. 174:341–352. http://dx.doi.org/10.1016/j.rse.2015.11.003.
Nagler, P., Glenn, E., Nguyen, U., Scott, R., Doody, T., 2013. Estimating riparian and agricultur-
al actual evapotranspiration by reference evapotranspiration and MODIS enhanced veg-
etation index. Remote Sens. 5:3849–3871. http://dx.doi.org/10.3390/rs5083849.
Oyler, J.W., Ballantyne, A., Jencso, K., Sweet, M., Running, S.W., 2015. Creating a
topoclimatic daily air temperature dataset for the conterminous United States using
homogenized station data and remotely sensed land skin temperature. Int.
J. Climatol. 35:2258–2279. http://dx.doi.org/10.1002/joc.4127.
Raymond, L.H., Rezin, K.V., 1989. Evapotranspiration estimates using remote-sensing
data, Parker and Palo Verde valleys, Arizona and California. USGS Water-Supply
Paper 2334 (p. 18).
Semmens, K.A., Anderson, M.C., Kustas, W.P., Gao, F., Alfieri, J.G., McKee, L., Prueger, J.H.,
Hain, C.R., Cammalleri, C., Yang, Y., Xia, T., Sanchez, L., Mar Alsina, M., Vélez, M.,
2015. Monitoring daily evapotranspiration over two California vineyards using
Landsat 8 in a multi-sensor data fusion approach. Remote Sens. Environ. http://dx.
doi.org/10.1016/j.rse.2015.10.025.
Senay, G.B., Bohms, S., Singh, R.K., Gowda, P.H., Velpuri, N.M., Alemu, H., Verdin, J.P., 2013.
Operational evapotranspiration mapping using remote sensing and weather
datasets: a new parameterization for the SSEB approach. J. Am. Water Res. Assoc.
49:577–591. http://dx.doi.org/10.1111/jawr.12057.
112 G.B. Senay et al. / Remote Sensing of Environment 202 (2017) 98–112
Senay, G.B., Velpuri, N.M., Bohms, S., Demissie, Y., Gebremichael, M., 2014. Understanding
the hydrologic sources and sinks in the Nile Basin using multisource climate and re-
mote sensing data sets. Water Resour. Res.:8625–8650 http://dx.doi.org/10.1002/
2013WR015231.
Senay, G.B., Friedrichs, M., Singh, R.K., Velpuri, N.M., 2016. Evaluating Landsat 8 evapo-
transpiration for water use mapping in the Colorado River Basin. Remote Sens. Envi-
ron. 185:171–185. http://dx.doi.org/10.1016/j.rse.2015.12.043.
Shelton, M.L., 1987. Irrigation induced change in vegetation and evapotranspiration in the
Central Valley of California. Landsc. Ecol. 1:95–105. http://dx.doi.org/10.1007/
BF00156231.
Singh, R.K., Liu, S., Tieszen, L.L., Suyker, A.E., Verma, S.B., 2011. Estimating seasonal evapo-
transpiration from temporal satellite images. Irrig. Sci. 30:303–313. http://dx.doi.org/
10.1007/s00271-011-0287-z.
Singh, R.K., Senay, G., Velpuri, N., Bohms, S., Scott, R., Verdin, J., 2014. Actual evapotrans-
piration (water use) assessment of the Colorado River Basin at the Landsat resolution
using the operational simplified surface energy balance model. Remote Sens. 6:
233–256. http://dx.doi.org/10.3390/rs6010233.
Sorooshian, S., AghaKouchak, A., Li, J.L., 2014. Influence of irrigation on land hydrological
processes over California. J. Geophys. Res.-Atmos. 119:13137–13152. http://dx.doi.
org/10.1002/2014jd022232.
Tulbure, M.G., Broich, M., Stehman, S.V., Kommareddy, A., 2016. Surface water extent dy-
namics from three decades of seasonally continuous Landsat time series at subconti-
nental scale in a semi-arid region. Remote Sens. Environ. 178:142–157. http://dx.doi.
org/10.1016/j.rse.2016.02.034.
Velpuri, N.M., Senay, G.B., 2013. Analysis of long-term trends (1950–2009) in precip-
itation, runoff and runoff coefficient in major urban watersheds in the United
States. Environ. Res. Lett. 8, 024020. http://dx.doi.org/10.1088/1748-9326/8/2/
024020.
Velpuri, N.M., Senay, G.B., Singh, R.K., Bohms, S., Verdin, J.P., 2013. A comprehensive eval-
uation of two MODIS evapotranspiration products over the conterminous United
States: using point and gridded FLUXNET and water balance ET. Remote Sens. Envi-
ron. 139:35–49. http://dx.doi.org/10.1016/j.rse.2013.07.013.
Westerling, A.L., Hidalgo, H.G., Cayan, D.R., Swetnam, T.W., 2006. Warming and earlier
spring increase Western U.S. forest wildfire activity. Science 313, 940–943.
Zhu, Z., Wang, S., Woodcock, C.E., 2015. Improvement and expansion of the Fmask algo-
rithm: cloud, cloud shadow, and snow detection for Landsats 4–7, 8, and Sentinel 2
images. Remote Sens. Environ. 159:269–277. http://dx.doi.org/10.1016/j.rse.2014.
12.014.